OpenAI and berries

Day 258 / 366

So we all know how ChatGPT was not able to solve a simple question — How many R’s are there in the word “Strawberry”?

Many speculated that this is a fundamental issue with LLMs since they work on tokens and not letters. One token might have multiple letters, so any question related to letter counting would be tough to solve with LLMs.

This is why the new O1 models were impressive, they were able to solve the strawberry question using the thinking step.

On looking more into it I found out that O1 uses something called a “chain-of-thought” prompting. Instead of coming up with an answer directly, this involves asking the LLM to come up with the steps to get the answer instead. By breaking down the problem into smaller steps, the chances of error reduces. This is not a new concept and was actually introduced in a paperback in 2022.

So O1 uses chain-of-thought to come up with steps to solve a problem and then uses AI agents to perform those steps. This is how it adds reasoning capabilities to the previously not-so-smart AI model.